Pulmonary Nodule Classification in CT Images Using ResNet18-Attention-LSTM Network47

Abstract

Pulmonary nodules serve as a common indicator for lung tumors, and prompt identification is essential in the early phase to prolong survival. Now, Computed Tomography (CT) is often used to check the lungs for lung cancer and determine if there is any. The old way of reading CT images manually is labor-intensive, time-consuming, and relies heavily on the experience of radiologists; therefore, different people may have different diagnoses. Recently, some deep learning methods have shown good results in medical image analysis and are now being used to automatically detect lung nodules. Due to the problem that traditional Convolutional Neural Networks (CNN) cannot use three-dimensional spatial information of pulmonary nodules well, this paper presents a pulmonary nodule CT image recognition model based on ResNet18, Attention Pooling and Long Short-Term Memory (LSTM) networks. Experiments will be carried out on the NoduleMNIST3D dataset, and the results of the model will be compared through CNN ablation studies, Recurrent Neural Network (RNN) ablation studies and attention mechanism comparison experiments. According to the findings, the integrated ResNet-Attention-LSTM framework has achieved a classification precision of 88.06% and an Area Under the Curve (AUC) of 0.9195 on the test set; it exceeds that of the baseline CNN and Convolutional RNN (CRNN) architectures. Based on these outcomes, it has been concluded that the merger of residual learning, attention mechanisms and temporal sequence models is suitable for improving the accuracy of pulmonary nodule classification and building intelligent lung cancer auxiliary diagnosis systems.